ANTHROPIC TOOL USE

A code sandbox for Claude.

Define a run_code tool in the Messages API and hand Claude a real interpreter — every call executes inside an isolated microVM, not on your host. Feed the result back and let Claude reason over the output.

Start free — $5 credits Coding agents guide

Works with the real anthropic SDK · no card · $5 free credits.

tool_use.py

from anthropic import Anthropic
from pilldrops import Sandbox

client = Anthropic()
tools = [{"name": "run_code", "description": "Run Python",
  "input_schema": {"type": "object",
    "properties": {"code": {"type": "string"}}}}]

msg = client.messages.create(model="claude-opus-4-5",
  max_tokens=1024, tools=tools,
  messages=[{"role": "user",
    "content": "Sum the primes under 100."}])

for block in msg.content:
  if block.type == "tool_use":
    with Sandbox.create(template="code-interpreter") as sb:
      r = sb.run(f"python -c '{block.input['code']}'")
      print(r.stdout)   # 1060

tool-use

ready out of the box

claude-terminal

CLI template

hosted

MCP server

24h

max session

free credits, no card

[01] — Tool use / sandbox exec

From tool call to executed result

Claude decides when to run code; PillDrops decides where. You register a run_code tool in the Messages API, and when Claude returns a tool_use block you execute the code inside a throwaway microVM — then send the stdout back as a tool_result so Claude can finish the answer.

STEP 01

Define the tool

Add a run_code entry to the tools array with an input schema for a code string. Claude now knows it can offload computation instead of guessing.

"name": "run_code",
"input_schema": { "code": "string" }

Messages APItool_use

STEP 02

Execute + feed back

On a tool_use block, spin up a code-interpreter sandbox, run the code, capture r.stdout, and append it as a tool_result. Claude reads real output — never a hallucinated number.

with Sandbox.create(...) as sb:
  r = sb.run(cmd)
  send(r.stdout)

code-interpreterisolated

Code interpreter guide →

[02] — Claude Code / in a microVM

Run Claude Code in an isolated box

Want the full Claude Code CLI, not just a function? The claude-terminal template boots a microVM with the Claude Code CLI installed and its own repo checkout — so the agent can edit files, run tests, and iterate without touching your machine. Give it real cores and a full day to work.

terminal

$ pilldrops sandbox create --template claude-terminal \
    --vcpu 4 --mem 8192 --timeout 86400
✓ sandbox sb_9f3c live · claude-terminal · 4 vCPU / 8 GB
✓ Claude Code CLI ready · repo checked out at /workspace
  session expires in 24h · pause/resume anytime

claude-terminal4 vCPU / 8 GB24h timeoutown repo checkout

[03] — Zero install / hosted MCP

Or just add the MCP

Skip the SDK plumbing entirely. Point Claude at the hosted PillDrops MCP server and it gets create, exec, file, and log tools over an OAuth-authed HTTP transport — nothing to install, no keys to paste.

add-mcp

$ claude mcp add --transport http pilldrops https://pilldrops.com/mcp
✓ added pilldrops (http) · OAuth · 73 tools available
  create_sandbox · run_command · write_file · read_file · get_logs …

Fourteen tools in total — create_sandbox, run_command, write_file, get_metrics, pause, resume, and more. Read the MCP server docs →

SHIP IT

Give Claude a place to run code

Real anthropic SDK, real pilldrops sandboxes. Wire up tool use, drop in the claude-terminal template, or add the hosted MCP — and start on $5 of free credits.

Start free Read the docs